toy-robust-backend-slam

• Goal: Toy and hands-on DCS codes for 2D graph SLAM to understand the importance of a robust back-end system for robust SLAM.
• Some explanations (pdf, Korean) can be found here.

Notes

• The almost parts of this repository's code are barrowed from https://github.com/mpkuse/toy-pose-graph-optimization-ceres.
• I added some useful other datasets from Luca Carlone's site.
• I refactored codes and directory structures for better understanding.
• I added my own results and discussions with repect to DCS On/Off, a portion of outliers, and datasets.
• I added code-block-explanation slides.
• (To do) I extended the original 2D example to 3D (i.e., SE3) examples.

How to run

• move to DCS-ceres directory just run (dependency: ceres is required)

$ cd DCS-ceres
$ ./do_build.sh DATASET_NAME_WITHOUGH_DOTG2O NUM_OUTLIER_LOOPS DSC_ON

• for example,

$ ./do_build.sh INTEL 50 1 # USING DCS, with 50 outlier loops on INTEL dataset
$ ./do_build.sh INTEL 50 0 # NOT USING DCS, with 50 outlier loops on INTEL dataset

What is DCS?

• As its name, literally, it dynamically scales the covariance.
• At the code level, it is more easy to understand (just 4 lines are added.)
• Simply, $\Phi$ is a free parameter (i.e., user parameter). In the author's [original paper ICRA 13), they changed this value from 0.1 to 10 and reported the performances for those ranges.

• If "psi" (denoted s in the original paper) is equal to 1, it is same as the original one, which is not robust to the outlier loops.

• Here, the detailed exaplanation of the mathematical derivation is ommitted and refer the original paper

Experimental results

dataset: INTEL

• These examples were acquired under default ceres solver option and $\Phi = 0.5$.
• As expected, without the robust back-end system such as DCS, the optimized graph topology easily collapsed.
• More exmaples are found in DCS-ceres/docs/{DATASET_NAME}.